A method, device, equipment and medium for detecting production conditions of pelletized ore

Through the optimization of image feature extraction model and text encoder, combined with multi-task classification model, the problem of low detection accuracy of pellet mine production conditions is solved, and more efficient detection accuracy and accuracy are achieved.

CN118072247BActive Publication Date: 2024-11-22湖南工商大学
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410193740.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-02-21
Publication Date
2024-11-22
Estimated Expiration
2044-02-21

AI Technical Summary

Technical Problem

The current pellet mine production conditions have low accuracy, and are affected by factors such as manual operation subjectivity, water mist, dust, vibration and uneven light, resulting in untimely regulation and low efficiency.

Method used

The image feature extraction model and text encoder are used for optimization, combined with the multi-task classification model, by obtaining the image features and text features of the production working condition image, the marginal probability is calculated to improve the detection accuracy.

Benefits of technology

It improves the accuracy and efficiency of pellet mine production conditions detection, reduces the amount of optimization data, improves the quality of image feature extraction models and text encoders, and ensures the accuracy of the multi-task classification model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118072247B_ABST
    Figure CN118072247B_ABST
Patent Text Reader

Abstract

The present application relates to the field of mineral industry detection technology, and provides a method, device, equipment and medium for detecting the production condition of pellet ore. The method includes: performing text annotation to obtain the label of the production condition image; using the image feature extraction model to extract features to obtain image features; using the text encoder to encode the label to obtain text features; optimizing the image feature extraction model and the text encoder; the optimized text encoder obtains the final text features; the optimized image feature extraction model obtains the final image features of the production sample image; the multi-task classification model obtains the marginal probability between the production sample image and the final text features; the multi-task classification model is optimized to obtain the final marginal probability between the production image to be detected and each final text feature; and the condition detection result is obtained based on the final marginal probability. The method of the present application can effectively improve the detection accuracy of the production condition detection of pellet ore.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of mineral industry detection technology, and in particular to a method, device, equipment and medium for detecting production conditions of pelletized ore. Background Art

[0002] At present, with the continuous development of computer technology, its technology has entered all walks of life, especially the rapid development of deep learning and large models, which has greatly improved the efficiency of various industries. Pelletized ore is an important raw material in contemporary blast furnace ironmaking. It has the advantages of high iron content, uniform particle size, good air permeability, and easy reduction. It can provide stable and high-quality raw materials for blast furnace smelting and achieve energy conservation and emission reduction before ironmaking.

[0003] At present, the detection and control of the working condition indicators of the pelletizing process mainly rely on manual work. Taking the pelletizing workshop of a domestic steel plant as an example, the operator estimates the quality characteristics of the pellets, including the particle size and brightness (humidity) change trend (commonly known as "watching the ball") by observing the pellet particles in the disc pelletizing machine. Based on this, the production conditions of the pellets are judged by experience and the amount of water added is manually adjusted to make the quality indicators of the pellets meet the standards. This production method of manual operation control is highly subjective, has large errors, low efficiency, and high labor intensity, resulting in delayed feedback of the measurement results, which makes the regulation untimely. Some researchers currently use machine vision methods for production measurement. Compared with traditional manual methods, the efficiency has been improved and the error has been reduced, but there are certain requirements for the data scale and model structure, and it needs to go through a complex image segmentation process. It is easily affected by factors such as water mist, dust, vibration, and uneven lighting at the production site, resulting in low accuracy in the detection of pellet production conditions. Summary of the invention

[0004] The present application provides a method, device, equipment and medium for detecting the working condition of pelletizing production, which can solve the problem of low accuracy in detecting the working condition of pelletizing production.

[0005] In a first aspect, an embodiment of the present application provides a method for detecting a pellet production condition, the method comprising:

[0006] Acquire multiple production working condition images of pelletizing ore, and perform text annotation on each production working condition image to obtain labels of the production working condition images;

[0007] The image feature extraction model is used to extract features from each production condition image to obtain image features of the production condition image;

[0008] The text encoder is used to encode the label of each production condition image to obtain the text features of the production condition image;

[0009] According to the image features and text features of all production working condition images, the image feature extraction model and the text encoder are optimized to obtain an optimized image feature extraction model and an optimized text encoder;

[0010] The optimized text encoder is used to obtain the final text features of each production condition image;

[0011] The final image features of the production sample image are obtained by using the optimized image feature extraction model; the production sample image is a production image of pellet ore with actual labels;

[0012] Based on the final image features of the production sample image and all the final text features, a multi-task classification model is used to obtain the marginal probability between the production sample image and each final text feature;

[0013] The multi-task classification model is optimized according to all marginal probabilities to obtain an optimized multi-task classification model, and the optimized multi-task classification model is used to obtain the final marginal probability between the production image of the pellet production that needs to be inspected for working conditions and each final text feature;

[0014] Based on all final marginal probabilities, the operating condition detection results of pellet production that requires operating condition detection are obtained.

[0015] Optionally, an image feature extraction model is used to extract features from each production condition image to obtain image features of the production condition image, including:

[0016] For each production condition image, use the image feature extraction model to perform the following steps:

[0017] Encoding the production working condition image to obtain the initial image features of the production working condition image;

[0018] Obtaining sequence features of initial image features, and converting the initial image features into sequence image features having the same dimension as the sequence features;

[0019] The image features of the production condition image are obtained by multiplying the sequence image features with the initial image features.

[0020] Optionally, the image feature extraction model and the text encoder are optimized according to the image features and text features of all production condition images to obtain an optimized image feature extraction model and an optimized text encoder, including:

[0021] Determine whether the proportion of all image features that achieve the expected quality is greater than the preset image proportion;

[0022] If the proportion of all image features that achieve the expected quality is greater than the preset image proportion, then determine whether the proportion of all text features that achieve the expected quality is greater than the preset text proportion;

[0023] If the proportion of all text features that reach the expected quality is greater than the preset text proportion, the image feature extraction model is used as the optimized image feature extraction model, and the text encoder is used as the optimized text encoder;

[0024] If the proportion of all text features that reach the expected quality is less than or equal to the preset text proportion, adjust the parameters in the text encoder, and return to the step of using the text encoder to encode the label of each production condition image respectively to obtain the text features of the production condition image;

[0025] If the proportion of all image features that meet the expected quality is less than or equal to the preset image proportion, the parameters in the image feature extraction model are adjusted, and the step of using the image feature extraction model to extract features from each production condition image to obtain the image features of the production condition image is returned.

[0026] Optionally, based on the final image features and all final text features of the production sample image, a multi-task classification model is used to obtain a marginal probability between the production sample image and each final text feature, including:

[0027] Calculate the similarity between the final image features of the production sample image and each final text feature using a multi-task classification model;

[0028] For each final text feature, the marginal probability between the production sample image and the final text feature is calculated according to the similarity between the final image feature of the production sample image and the final text feature.

[0029] Optionally, a multi-task classification model is used to calculate the similarity between the final image features of the production sample image and each final text feature, including:

[0030] By formula:

[0031]

[0032] Calculate the similarity logit(c,s,d|x) between the production sample image y and the xth final text feature;

[0033] Wherein, c represents the c-class label, s represents the s-class label, d represents the d-class label, the label corresponding to the x-th final text feature belongs to one of the c-class label, the s-class label or the d-class label, x=1,2,...,X, X represents the number of production condition images, F(y) represents the final image feature of the production sample image y, and G(x) represents the x-th final text feature;

[0034] According to the similarity between the final image feature of the production sample image and the final text feature, the marginal probability between the production sample image and the final text feature is calculated, including:

[0035] By formula:

[0036]

[0037] Calculate the marginal probability between the production sample image y and the xth final text feature

[0038] Among them, τ represents a parameter.

[0039] Optionally, optimize the multi-task classification model based on all marginal probabilities, including:

[0040] By formula:

[0041] L Totle =w 1 L s (y)+w 2 L c (y)+w 3 L d (y)

[0042] Calculate the loss function value L of the multi-task classification model Totle ;

[0043] Among them, w 1 、w 2 and w 3 All represent parameters, L s (y) represents the loss function between the predicted label of the production sample image y and the class label s, L c (y) represents the loss function between the predicted label of the production sample image y and the c-class label, L d (y) represents the loss function between the predicted label of the production sample image y and the d-class label:

[0044]

[0045] Where p(h|y) represents the relationship between the production sample image y and the label h. When the label of the production sample image y is the same as the label h, p(h|y)=1. When the label of the production sample image y is different from the label h, p(h|y)=0. S represents the set of s-type labels. When the label corresponding to the final text feature is label h, the marginal probability between the production sample image y and the final text feature, p(j|y) represents the relationship between the production sample image y and label j, when the label of the production sample image y is the same as label j, p(j|y)=1, when the label of the production sample image y is different from label j, p(j|y)=0, C represents the set of c-type labels, When the label corresponding to the final text feature is label j, the marginal probability between the production sample image y and the final text feature is expressed. p(m|y) represents the relationship between the production sample image y and label m. When the label of the production sample image y is the same as label m, p(m|y)=1. When the label of the production sample image y is different from label m, p(m|y)=0. D represents the set of d-class labels. Indicates the marginal probability between the generated sample image y and the final text feature when the label corresponding to the final text feature is label m;

[0046] Determine whether the loss function value of the multi-task classification model is less than the preset loss function value;

[0047] If yes, the multi-task classification model is used as the optimized multi-task classification model;

[0048] Otherwise, adjust the parameters in the multi-task classification model, return the final image features and all the final text features based on the production sample image, and use the multi-task classification model to obtain the marginal probability between the production sample image and each final text feature.

[0049] Optionally, based on all final marginal probabilities, the operating condition test results of pellet production that require operating condition test are obtained, including:

[0050] Select the final marginal probability with the largest value from all final marginal probabilities as the target marginal probability;

[0051] The label corresponding to the target marginal probability is used as the operating condition detection result of the pellet production that needs to be tested.

[0052] In a second aspect, an embodiment of the present application provides a pelletizing production condition detection device, comprising:

[0053] The labeling module obtains multiple production working condition images of pelletizing ore, and performs text labeling on each production working condition image to obtain the label of the production working condition image;

[0054] The feature extraction module uses the image feature extraction model to extract features from each production condition image to obtain image features of the production condition image;

[0055] The encoding module uses a text encoder to encode the label of each production condition image to obtain the text features of the production condition image;

[0056] The first optimization module optimizes the image feature extraction model and the text encoder according to the image features and text features of all production working condition images to obtain an optimized image feature extraction model and an optimized text encoder;

[0057] The first acquisition module uses the optimized text encoder to obtain the final text features of each production condition image;

[0058] The second acquisition module uses the optimized image feature extraction model to acquire the final image features of the production sample image; the production sample image is a production image of the pellet ore with actual labels;

[0059] A marginal probability acquisition module, based on the final image features of the production sample image and all the final text features, uses a multi-task classification model to acquire the marginal probability between the production sample image and each final text feature;

[0060] The second optimization module optimizes the multi-task classification model according to all marginal probabilities to obtain an optimized multi-task classification model, and uses the optimized multi-task classification model to obtain the final marginal probability between the production image of the pellet production that needs to be inspected for working conditions and each final text feature;

[0061] The detection module obtains the working condition detection results of the pellet production that needs to be detected based on all final marginal probabilities.

[0062] In a third aspect, an embodiment of the present application provides a terminal device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned pelletizing ore production condition detection method when executing the above-mentioned computer program.

[0063] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the above-mentioned pelletizing production condition detection method is implemented.

[0064] The above solution of the present application has the following beneficial effects:

[0065] In an embodiment of the present application, a plurality of production condition images are obtained, and text annotation is performed on each production condition image respectively to obtain labels of the production condition images, and then an image feature extraction model is used to extract features of each production condition image respectively to obtain image features of the production condition images, and then a text encoder is used to encode the labels of each production condition image respectively to obtain text features of the production condition images, and then the image feature extraction model and the text encoder are optimized according to the image features and text features of all production condition images to obtain an optimized image feature extraction model and an optimized text encoder, and then the optimized text encoder is used to obtain the text features of each production condition image. The final text features of the image are obtained, and then the optimized image feature extraction model is used to obtain the final image features of the production sample image. Based on the final image features and all the final text features of the production sample image, the multi-task classification model is used to obtain the marginal probability between the production sample image and each final text feature. Then, the multi-task classification model is optimized according to all the marginal probabilities to obtain the optimized multi-task classification model. The optimized multi-task classification model is used to obtain the final marginal probability between the production image of the pellet production that needs to be tested for working conditions and each final text feature. Finally, based on all the final marginal probabilities, the working condition detection results of the pellet production that needs to be tested for working conditions are obtained. Among them, only the image features and text features of multiple production condition images are used to optimize the image feature extraction model and the text encoder, which reduces the amount of data involved in the optimization and improves the optimization efficiency. At the same time, the quality of the image feature extraction model and the text encoder is improved. The multi-task classification model is optimized according to all marginal probabilities, which can improve the quality of the multi-task classification model. The final marginal probability obtained by using a high-quality multi-task classification model has high accuracy. The working condition detection results of pellet production that require working condition detection are obtained based on the final marginal probability with high accuracy, which can effectively improve the detection accuracy of pellet production working condition detection.

[0066] Other beneficial effects of the present application will be described in detail in the subsequent specific implementation section. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0068] Figure 1 A flow chart of a pelletizing production condition detection method provided in one embodiment of the present application;

[0069] Figure 2A schematic diagram of the structure of an improved Unet provided in one embodiment of the present application;

[0070] Figure 3 A schematic diagram of a specific process of a pellet production condition detection method provided in an embodiment of the present application;

[0071] Figure 4 A schematic diagram of the working condition detection results of pellet production that requires working condition detection provided in one embodiment of the present application;

[0072] Figure 5 A schematic diagram of the structure of a pellet production condition detection device provided in one embodiment of the present application;

[0073] Figure 6 A schematic diagram of the structure of a terminal device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0074] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0075] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0076] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0077] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0078] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0079] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0080] In view of the problem of low accuracy of existing pelletizing production condition detection, an embodiment of the present application provides a pelletizing production condition detection method. In the embodiment of the present application, the pelletizing production condition detection method obtains multiple production condition images, and respectively annotates each production condition image with text to obtain the label of the production condition image, then uses an image feature extraction model to extract features from each production condition image to obtain image features of the production condition image, and then uses a text encoder to encode the label of each production condition image to obtain text features of the production condition image, and then optimizes the image feature extraction model and the text encoder according to the image features and text features of all production condition images to obtain an optimized image feature extraction model and an optimized text encoder. The optimized text encoder is then used to obtain the final text features of each production condition image, and then the optimized image feature extraction model is used to obtain the final image features of the production sample images, and then based on the final image features of the production sample images and all the final text features, the multi-task classification model is used to obtain the marginal probability between the production sample images and each final text feature, and then the multi-task classification model is optimized according to all the marginal probabilities to obtain the optimized multi-task classification model, and the optimized multi-task classification model is used to obtain the final marginal probability between the production images of the pellet production that needs to be tested for working conditions and each final text feature, and finally the working condition detection results of the pellet production that needs to be tested for working conditions are obtained based on all the final marginal probabilities. Among them, only the image features and text features of multiple production condition images are used to optimize the image feature extraction model and the text encoder, which reduces the amount of data involved in the optimization and improves the optimization efficiency. At the same time, the quality of the image feature extraction model and the text encoder is improved. The multi-task classification model is optimized according to all marginal probabilities, which can improve the quality of the multi-task classification model. The final marginal probability obtained by using a high-quality multi-task classification model has high accuracy. The working condition detection results of pellet production that require working condition detection are obtained based on the final marginal probability with high accuracy, which can effectively improve the detection accuracy of pellet production working condition detection.

[0081] Next, the pelletizing production condition detection method provided in this application is exemplified.

[0082] like Figure 1 As shown, the pellet production condition detection method provided in this application includes the following steps:

[0083] Step 11, obtaining multiple production working condition images of pelletizing ore, and respectively annotating each production working condition image with text to obtain a label of the production working condition image.

[0084] It should be noted that the above production condition images are production images when pellets are being processed.

[0085] Exemplarily, the label can be a label of multiple types, such as particle size type (powdery, large particle size, small particle size, moderate particle size), processing environment conditions (dust interference, fog interference, insufficient light, normal, less water, more water), particle distribution conditions (uneven, uniform), etc. For a production condition image, the production condition image can be annotated with text manually or by computer software such as labelme and CVAT to obtain the label of the production condition image, such as the label of the production condition image is powdery. For the production condition of pellets, the labels of particle size type and particle distribution can describe the particle state of pellets, and the label of processing environment conditions can describe the processing environment when processing pellets, such as: because the water body affects the friction of the surface of the sphere, the smooth sphere reflects the light, causing the brightness of the production image to be higher than the normal brightness, which means that the water volume is less.

[0086] It is worth mentioning that by performing text annotation on the production condition images, multiple production condition images with labels are obtained, which is convenient for subsequent calculations.

[0087] Step 12: Use the image feature extraction model to extract features from each production condition image to obtain image features of the production condition image.

[0088] In some embodiments of the present application, the above-mentioned step of using the image feature extraction model to extract features from each production condition image to obtain image features of the production condition image is specifically as follows:

[0089] It should be noted that the image feature extraction model includes an image encoding module and an image sequence feature extraction module.

[0090] For each production condition image, use the image feature extraction model to perform the following steps:

[0091] The first step is to encode the production condition image to obtain the initial image features of the production condition image.

[0092] Specifically, the image encoding module in the image feature extraction model is used to encode the production condition image to obtain the initial image features of the production condition image.

[0093] The second step is to obtain the sequence features of the initial image features and convert the initial image features into sequence image features with the same dimension as the sequence features.

[0094] Specifically, the image sequence feature extraction module is used to obtain the sequence features of the initial image features, and then the reshape function is used to convert the initial image features into sequence image features with the same dimension as the sequence features.

[0095] The third step is to multiply the sequence image features with the initial image features to obtain the image features of the production condition image.

[0096] Exemplarily, the image sequence feature extraction module may be an improved algorithm ViT-32 of a pre-trained vision transformer (ViT, VisionTransformer), and the image encoding module may be a pre-trained improved U-shaped semantic segmentation network (Unet), all 3×3 convolutions of the improved Unet are multi-dimensional fusion modules, and a channel and space dual attention mechanism is added at the jump connection of each layer. The above-mentioned multi-dimensional fusion module includes three sequentially connected feature extraction units and an addition unit, the input end of the first feature extraction unit is the input end of the multi-dimensional fusion module, the output end of each feature extraction unit is connected to the input end of the addition unit, and the output end of the addition unit is the output end of the multi-dimensional fusion module. The above-mentioned feature extraction unit includes a convolution subunit and an activation function subunit connected in sequence, the input end of the convolution subunit is the input end of the feature extraction unit, the output end of the activation function subunit is the output end of the feature extraction unit, the convolution in the convolution subunit is a 3×3 convolution, and the activation function subunit is a ReLU and Drop function.

[0097] It is worth mentioning that replacing all 3×3 convolutions of Unet with multi-dimensional fusion modules can improve the information accuracy of the acquired initial image features. The image features obtained based on the high-precision initial image features have high accuracy in representing the information.

[0098] The above-mentioned image encoding module is exemplarily described below with reference to a specific example.

[0099] like Figure 2 As shown, Figure 2 (a) is an improved Unet. MDA-Conv means replacing the 3×3 convolution with the encoder of the multi-dimensional fusion module. A means the channel and spatial dual attention mechanism. The decoder is the decoder of the traditional Unet. Up means upsampling. Figure 2(b) is the structure of the multi-dimensional fusion module, where 3×3 represents the convolution subunit of 3×3 convolution, ReLU+Drop represents the activation function subunit, and + represents the addition unit.

[0100] Step 13, using a text encoder to encode the label of each production condition image respectively to obtain the text features of the production condition image.

[0101] In some embodiments of the present application, the text encoder may be an improved model of a generative pre-trained transformer (GPT, Generative Pre Trained transformer) GPT-2, and GPT-2 is pre-trained.

[0102] For example, the pre-trained GPT-2 can be run using computer software such as MATLAB to obtain text features of each production condition image.

[0103] It is worth mentioning that data with labels in text form are not convenient for direct calculation with data from other modalities. Acquiring text features of production condition images can convert text data into a data form that is easy to calculate while representing the information of the labels of production condition images.

[0104] Step 14, optimizing the image feature extraction model and the text encoder according to the image features and text features of all production working condition images to obtain an optimized image feature extraction model and an optimized text encoder.

[0105] Specifically, it is determined whether the proportion of all image features that achieve the expected quality is greater than a preset image proportion.

[0106] If the proportion of all image features that achieve the expected quality is greater than the preset image proportion, then it is determined whether the proportion of all text features that achieve the expected quality is greater than the preset text proportion.

[0107] If the proportion of all text features that achieve the expected quality is greater than the preset text proportion, the image feature extraction model is used as the optimized image feature extraction model, and the text encoder is used as the optimized text encoder.

[0108] If the proportion of all text features that meet the expected quality is less than or equal to the preset text proportion, the parameters in the text encoder are adjusted, and the step of using the text encoder to encode the label of each production condition image separately to obtain the text features of the production condition image is returned.

[0109] If the proportion of all image features that meet the expected quality is less than or equal to the preset image proportion, the parameters in the image feature extraction model are adjusted, and the step of using the image feature extraction model to extract features from each production condition image to obtain the image features of the production condition image is returned.

[0110] For example, for 10 production condition images, for 10 image features and 10 text features, 7 image features and 4 text features meet the expected quality, reaching 60% of the preset image proportion, and not reaching 60% of the preset text proportion. This indicates that at this time the performance of the image feature extraction module meets expectations, while the performance of the text encoder does not meet expectations. The parameters in the text encoder are adjusted, and 10 text features are re-acquired. At this time, 8 text features meet the expected quality, indicating that at this time the performance of the text encoder meets expectations and is an optimized text encoder.

[0111] It is worth mentioning that in this step, only a small amount of data is needed to optimize the text encoder and image feature extraction model, reduce the time consumption of the text encoder and image feature extraction model in processing data, and improve the optimization efficiency.

[0112] Step 15, using the optimized text encoder to obtain the final text features of each production condition image.

[0113] Specifically, the optimized text encoder is used to encode the label of each production condition image respectively to obtain the final text features of the production condition image.

[0114] For example, the optimized text encoder may be run by computer software such as MATLAB to obtain the final text features of the production condition image.

[0115] It is worth mentioning that the final text features obtained using the text encoder with expected performance are highly accurate.

[0116] Step 16, using the optimized image feature extraction model to obtain the final image features of the production sample image.

[0117] The production sample image is a production image of pelletized ore with actual labels. The production sample image can be the production condition image with labels obtained after text annotation as described above.

[0118] Specifically, the production sample image is encoded using the image encoding module in the optimized image feature extraction model to obtain the initial image features of the production sample image, and then the sequence features of the initial image features are obtained using the image sequence feature extraction module in the optimized image feature extraction model, and the initial image features are converted into sequence image features with the same dimension as the sequence features, and then the sequence image features are multiplied by the initial image features to obtain the final image features of the production sample image.

[0119] Exemplarily, the optimized image feature extraction model can be run using computer software such as MATLAB to obtain the final image features of the production sample image.

[0120] It is worth mentioning that the accuracy of the final image features of the production sample images obtained using the image feature extraction model with the expected performance is high.

[0121] Step 17, based on the final image features of the production sample image and all the final text features, a multi-task classification model is used to obtain the marginal probability between the production sample image and each final text feature.

[0122] In some embodiments of the present application, the multi-task classification model may be a multi-label decision tree model (ML-DT, Multi Label DecisionTree), and the step of obtaining the marginal probability between the production sample image and each final text feature using the multi-task classification model based on the final image feature and all final text features of the production sample image specifically includes:

[0123] In the first step, a multi-task classification model is used to calculate the similarity between the final image features of the production sample image and each final text feature.

[0124] Specifically, through the formula:

[0125]

[0126] Calculate the similarity logit(c,s,d|x) between the production sample image y and the xth final text feature.

[0127] Wherein, c represents the c-class label, s represents the s-class label, d represents the d-class label, the label corresponding to the x-th final text feature belongs to one of the c-class label, the s-class label or the d-class label, x=1,2,...,X, X represents the number of production condition images, F(y) represents the final image feature of the production sample image y, and G(x) represents the x-th final text feature.

[0128] For example, the c-type label is the above-mentioned particle size type (powdery, large particle size, small particle size, medium particle size), the s-type label is the above-mentioned processing environment condition (dust interference, fog interference, insufficient light, normal, less water, more water), and the d-type label is the above-mentioned particle distribution condition (uneven, even). If the label corresponding to the x-th final text feature is the d-type label, the calculated similarity between the production sample image y and the x-th final text feature is logit(d|x).

[0129] In the second step, for each final text feature, the marginal probability between the production sample image and the final text feature is calculated according to the similarity between the final image feature of the production sample image and the final text feature.

[0130] Specifically, through the formula:

[0131]

[0132] Calculate the marginal probability between the production sample image y and the xth final text feature

[0133] Among them, τ represents a parameter.

[0134] It can be understood that the above formula for calculating similarity logit(c,s,d|x) and calculating marginal probability It can be understood as an expression of a multi-task classification model.

[0135] Exemplarily, the label corresponding to the xth final text feature is a d-class label, then the calculated marginal probability between the production sample image y and the xth final text feature is

[0136] It is worth mentioning that by calculating the marginal probability of each final text feature and the production sample image, the degree of correlation between the production sample image and the final text feature can be reflected. The higher the marginal probability, the more similar the label of the production sample image is to the label corresponding to the final text feature.

[0137] The above steps are exemplarily described below with reference to a specific example.

[0138] like Figure 3As shown, the label of the production condition image enters the text encoder to obtain text features, the production condition image enters the image encoder module of the image feature extraction model to obtain initial image features, the initial image features are processed by the image sequence feature extraction module of the image feature extraction model to obtain sequence features, the initial image features are converted into sequence image features with the same dimension as the sequence features, the sequence image features are multiplied by the initial image features to obtain image features, the image features and text features are input into the multi-task classification model, and the marginal probability is output.

[0139] Step 18, optimizing the multi-task classification model according to all marginal probabilities to obtain an optimized multi-task classification model, and using the optimized multi-task classification model to obtain the final marginal probability between the production image of the pellet production that needs to be inspected for working conditions and each final text feature.

[0140] In some embodiments of the present application, the above-mentioned step of optimizing the multi-task classification model according to all marginal probabilities to obtain the optimized multi-task classification model, and using the optimized multi-task classification model to obtain the final marginal probability between the production image of pellet production that needs to be tested for working conditions and each final text feature specifically includes:

[0141] The first step is through the formula:

[0142] L Totle =w 1 L s (y)+w 2 L c (y)+w 3 L d (y)

[0143] Calculate the loss function value L of the multi-task classification model Totl e.

[0144] Among them, w 1 、w 2 and w 3 All represent parameters, L s (y) represents the loss function between the predicted label of the production sample image y and the class label s, L c (y) represents the loss function between the predicted label of the production sample image y and the c-class label, L d (y) represents the loss function between the predicted label of the production sample image y and the d-class label:

[0145]

[0146] Where p(h|y) represents the relationship between the production sample image y and the label h. When the label of the production sample image y is the same as the label h, p(h|y)=1. When the label of the production sample image y is different from the label h, p(h|y)=0. S represents the set of s-type labels. When the label corresponding to the final text feature is label h, the marginal probability between the production sample image y and the final text feature, p(j|y) represents the relationship between the production sample image y and label j, when the label of the production sample image y is the same as label j, p(j|y)=1, when the label of the production sample image y is different from label j, p(j|y)=0, C represents the set of c-type labels, When the label corresponding to the final text feature is label j, the marginal probability between the production sample image y and the final text feature is expressed. p(m|y) represents the relationship between the production sample image y and label m. When the label of the production sample image y is the same as label m, p(m|y)=1. When the label of the production sample image y is different from label m, p(m|y)=0. D represents the set of d-class labels. Indicates the marginal probability between the generated sample image y and the final text feature when the label corresponding to the final text feature is label m.

[0147] The second step is to determine whether the loss function value of the multi-task classification model is less than the preset value of the loss function.

[0148] If so, the multi-task classification model is taken as the optimized multi-task classification model.

[0149] Otherwise, adjust the parameters in the multi-task classification model, return the final image features and all the final text features based on the production sample image, and use the multi-task classification model to obtain the marginal probability between the production sample image and each final text feature.

[0150] In the third step, the optimized multi-task classification model is used to obtain the final marginal probability between the production image of the pellet production that needs to be inspected for working conditions and each final text feature.

[0151] Specifically, the optimized image feature extraction model is first used to obtain the final image features of the production images of pellet production that require working condition detection, and then the optimized multi-task classification model is used to obtain the final marginal probability between the final image features of the production images of pellet production that require working condition detection and each final text feature.

[0152] Exemplarily, the loss function value of the multi-task classification model is 3, which is greater than the preset value of the loss function of 2.7, indicating that the performance of the multi-task classification model at this time does not meet expectations. The parameters in the multi-task classification model are adjusted, the marginal probability between the production sample image and each final text feature is re-obtained, and the loss function value is calculated. The loss function value at this time is 1.8, indicating that the performance of the multi-task classification model at this time meets expectations and is an optimized multi-task classification model.

[0153] It should be noted that the parameter w in the above formula for calculating the loss function value is 1 、w 2 and w 3 , are set according to the actual situation of the task. For example, if the particle size detection type is the main task in a detection task, the parameter w is adjusted. 1 、w 2 and w 3 , making the loss function value more inclined to evaluate the accuracy of detecting particle size types.

[0154] It is worth mentioning that using the loss function value to optimize the multi-task classification model can improve the quality of the multi-task classification model and make the performance of the multi-task classification model reach the expected performance. At the same time, adjusting the parameters in the calculation formula of the loss function value according to the actual task requirements can make the multi-task classification model meet the task requirements and improve the rationality of the multi-task classification model.

[0155] Step 19, based on all final marginal probabilities, obtain the operating condition detection results of the pellet production that needs to be detected.

[0156] It should be noted that the operating condition detection is to detect the operating conditions of the pellet production that needs to be detected, and the above-mentioned operating condition detection results are labels (such as powdery, large particle size, small particle size, moderate particle size, fog, uneven, uniform, etc.).

[0157] In some embodiments of the present application, the above-mentioned step of obtaining the working condition detection result of pellet production that needs to be tested for working condition based on all final marginal probabilities is specifically as follows:

[0158] In the first step, the final marginal probability with the largest value is selected from all the final marginal probabilities as the target marginal probability.

[0159] In the second step, the label corresponding to the target marginal probability is used as the working condition detection result of the pellet production that needs to be tested.

[0160] Exemplarily, the final marginal probabilities are 2, 1.3, 4.2, 3.7, and 0.7, respectively. The final marginal probability of 4.2 is used as the target marginal probability. The label corresponding to the target marginal probability is fog. The operating condition detection result of the pellet production that needs to be tested is fog.

[0161] It should be noted that, for a production image of pellet production that is large in size or has complex working conditions and requires working condition detection, the production image is cropped into multiple production sub-images, and the working condition detection result of each production sub-image is obtained.

[0162] The following is an illustrative description of the working condition detection results of the pellet production that requires working condition detection, with reference to a specific example.

[0163] like Figure 4 As shown in the figure, the first row of the figure contains four production images of pellet production that need to be tested for working conditions, with the label type being particle size type, and the labels are powdery, large particle size, small particle size, and moderate particle size respectively; the second row of the figure contains four production images of pellet production that need to be tested for working conditions, with the label type being processing environment conditions, and the labels are dust interference, fog interference, insufficient light, and normal respectively; the third row of the figure contains two production images of pellet production that need to be tested for working conditions, with the label type being particle distribution conditions, and the labels are uneven and uniform respectively.

[0164] It is worth mentioning that only the image features and text features of multiple production condition images are used to optimize the image feature extraction model and the text encoder, which reduces the amount of data involved in the optimization and improves the optimization efficiency. At the same time, the quality of the image feature extraction model and the text encoder is improved. Optimizing the multi-task classification model according to all marginal probabilities can improve the quality of the multi-task classification model. The final marginal probability obtained by using a high-quality multi-task classification model has high accuracy. The condition detection results of pellet production that require condition detection are obtained based on the final marginal probability with high accuracy, which can effectively improve the detection accuracy of pellet production condition detection.

[0165] The pelletizing production condition detection device of the present application is exemplarily described below with reference to a specific example.

[0166] like Figure 5 As shown, the embodiment of the present application provides a pelletizing production condition detection device, and the pelletizing production condition detection device 500 includes:

[0167] The labeling module 501 obtains multiple production working condition images of pelletizing ore, and performs text labeling on each production working condition image to obtain a label of the production working condition image;

[0168] A feature extraction module 502 uses an image feature extraction model to extract features from each production condition image to obtain image features of the production condition image;

[0169] The encoding module 503 uses a text encoder to encode the label of each production condition image to obtain the text features of the production condition image;

[0170] A first optimization module 504 optimizes the image feature extraction model and the text encoder according to the image features and text features of all production working condition images to obtain an optimized image feature extraction model and an optimized text encoder;

[0171] A first acquisition module 505 uses an optimized text encoder to acquire final text features of each production condition image;

[0172] The second acquisition module 506 uses the optimized image feature extraction model to acquire the final image features of the production sample image; the production sample image is a production image of the pellet ore with actual labels;

[0173] A marginal probability acquisition module 507, based on the final image features of the production sample image and all the final text features, uses a multi-task classification model to acquire a marginal probability between the production sample image and each final text feature;

[0174] The second optimization module 508 optimizes the multi-task classification model according to all marginal probabilities to obtain an optimized multi-task classification model, and uses the optimized multi-task classification model to obtain a final marginal probability between a production image of pellet production that needs to be inspected for working conditions and each final text feature;

[0175] The detection module 509 obtains the working condition detection result of the pellet production that needs to be detected based on all the final marginal probabilities.

[0176] It should be noted that the information interaction, execution process, etc. between the above-mentioned devices / units are based on the same concept as the method embodiment of the present application. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0177] The technicians in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0178] like Figure 6 As shown, an embodiment of the present application provides a terminal device. The terminal device D10 of this embodiment includes: at least one processor D100 ( Figure 6 Only one processor is shown in the figure), a memory D101, and a computer program D102 stored in the memory D101 and executable on the at least one processor D100, wherein the processor D100 implements the steps of any of the above-mentioned method embodiments when executing the computer program D102.

[0179] Specifically, when the processor D100 executes the computer program D102, in an embodiment of the present application, a plurality of production working condition images are obtained, and text annotation is performed on each production working condition image respectively to obtain a label of the production working condition image, and then an image feature extraction model is used to extract features of each production working condition image respectively to obtain image features of the production working condition image, and then a text encoder is used to encode the label of each production working condition image respectively to obtain text features of the production working condition image, and then the image feature extraction model and the text encoder are optimized according to the image features and text features of all the production working condition images to obtain an optimized image feature extraction model and an optimized text encoder, and then the optimized image feature extraction model and the optimized text encoder are used. The text encoder after optimization obtains the final text features of each production working condition image, and then uses the optimized image feature extraction model to obtain the final image features of the production sample image, and then based on the final image features of the production sample image and all the final text features, uses the multi-task classification model to obtain the marginal probability between the production sample image and each final text feature, and then optimizes the multi-task classification model according to all the marginal probabilities to obtain the optimized multi-task classification model, and uses the optimized multi-task classification model to obtain the final marginal probability between the production image of the pellet production that needs to be tested for working condition and each final text feature, and finally obtains the working condition detection result of the pellet production that needs to be tested for working condition based on all the final marginal probabilities. Among them, only the image features and text features of multiple production condition images are used to optimize the image feature extraction model and the text encoder, which reduces the amount of data involved in the optimization and improves the optimization efficiency. At the same time, the quality of the image feature extraction model and the text encoder is improved. The multi-task classification model is optimized according to all marginal probabilities, which can improve the quality of the multi-task classification model. The final marginal probability obtained by using a high-quality multi-task classification model has high accuracy. The working condition detection results of pellet production that require working condition detection are obtained based on the final marginal probability with high accuracy, which can effectively improve the detection accuracy of pellet production working condition detection.

[0180] The processor D100 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0181] In some embodiments, the memory D101 may be an internal storage unit of the terminal device D10, such as a hard disk or memory of the terminal device D10. In other embodiments, the memory D101 may also be an external storage device of the terminal device D10, such as a plug-in hard disk, a smart memory card (SMC, SmartMedia Card), a secure digital (SD, Secure Digital) card, a flash card (Flash Card), etc. equipped on the terminal device D10. Further, the memory D101 may also include both an internal storage unit of the terminal device D10 and an external storage device. The memory D101 is used to store an operating system, an application program, a boot loader (BootLoader), data and other programs, such as the program code of the computer program, etc. The memory D101 may also be used to temporarily store data that has been output or is to be output.

[0182] An embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps in the above-mentioned method embodiments can be implemented.

[0183] An embodiment of the present application provides a computer program product. When the computer program product runs on a terminal device, the terminal device can implement the steps in the above-mentioned method embodiments when executing the computer program product.

[0184] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may at least include: any entity or device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium that can carry the computer program code to the pellet production working condition detection method device / terminal device. For example, a USB flash drive, a mobile hard disk, a magnetic disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0185] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0186] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0187] The above is a preferred embodiment of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles described in the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. A method for detecting pellet production conditions, characterized in that: include: Acquire multiple production working condition images of pelletizing ore, and respectively annotate each of the production working condition images with text to obtain labels of the production working condition images; Using an image feature extraction model to extract features from each of the production working condition images, to obtain image features of the production working condition images; Using a text encoder to encode the label of each of the production working condition images respectively to obtain text features of the production working condition images; According to the image features and text features of all production working condition images, the image feature extraction model and the text encoder are optimized to obtain an optimized image feature extraction model and an optimized text encoder; Using the optimized text encoder to obtain the final text features of each of the production working condition images; Acquiring final image features of the production sample image using the optimized image feature extraction model; The production sample image is a production image of pellets with actual labels; Based on the final image features and all the final text features of the production sample image, using a multi-task classification model to obtain a marginal probability between the production sample image and each of the final text features; The multi-task classification model is optimized according to all marginal probabilities to obtain an optimized multi-task classification model, and the optimized multi-task classification model is used to obtain the final marginal probability between the production image of the pellet production that needs to be tested for working conditions and each of the final text features; The operating condition detection result of the pellet production that needs to be detected is obtained based on all final marginal probabilities.

2. The pellet production condition detection method according to claim 1, characterized in that: The method of using the image feature extraction model to extract features from each production condition image to obtain image features of the production condition image includes: For each production condition image, use the image feature extraction model to perform the following steps: Encoding the production working condition image to obtain initial image features of the production working condition image; Acquire sequence features of the initial image features, and convert the initial image features into sequence image features having the same dimension as the sequence features; The sequence image feature is multiplied by the initial image feature to obtain the image feature of the production condition image.

3. The pellet production condition detection method according to claim 1, characterized in that: The image feature extraction model and the text encoder are optimized according to the image features and text features of all production working condition images to obtain the optimized image feature extraction model and the optimized text encoder, including: Determine whether the proportion of all image features that achieve the expected quality is greater than the preset image proportion; If the proportion of all image features that achieve the expected quality is greater than the preset image proportion, then determine whether the proportion of all text features that achieve the expected quality is greater than the preset text proportion; If the proportion of all text features that reach the expected quality is greater than the preset text proportion, the image feature extraction model is used as the optimized image feature extraction model, and the text encoder is used as the optimized text encoder; If the proportion of all text features that reach the expected quality is less than or equal to the preset text proportion, the parameters in the text encoder are adjusted, and the step of using the text encoder to encode the label of each of the production working condition images respectively to obtain the text features of the production working condition images is returned; If the proportion of all image features that meet the expected quality is less than or equal to the preset image proportion, the parameters in the image feature extraction model are adjusted, and the step of using the image feature extraction model to extract features from each of the production condition images to obtain the image features of the production condition images is returned.

4. The pellet production condition detection method according to claim 1, characterized in that: The method of obtaining the marginal probability between the production sample image and each of the final text features by using a multi-task classification model based on the final image features and all the final text features of the production sample image includes: Calculating the similarity between the final image feature of the production sample image and each of the final text features using the multi-task classification model; For each of the final text features, the marginal probability between the production sample image and the final text feature is calculated according to the similarity between the final image feature of the production sample image and the final text feature.

5. The pellet production condition detection method according to claim 4, characterized in that: The calculating the similarity between the final image feature of the production sample image and each of the final text features by using the multi-task classification model includes: By formula: Calculate the similarity logit(c,s,d|x) between the production sample image y and the xth final text feature; Wherein, c represents a c-class label, s represents an s-class label, d represents a d-class label, the label corresponding to the x-th final text feature belongs to one of the c-class label, the s-class label or the d-class label, x=1, 2, ..., X, X represents the number of the production working condition images, F(y) represents the final image feature of the production sample image y, and G(x) represents the x-th final text feature; The calculating the marginal probability between the production sample image and the final text feature according to the similarity between the final image feature of the production sample image and the final text feature comprises: By formula: Calculate the marginal probability between the production sample image y and the xth final text feature Among them, τ represents a parameter.

6. The pellet production condition detection method according to claim 1, characterized in that: Optimizing the multi-task classification model according to all marginal probabilities includes: By formula: L Totle =w1L s (y)+w2L c (y)+w3L d (and) Calculate the loss function value L of the multi-task classification model Totle ; Among them, w1, w2 and w3 are parameters, L s (y) represents the loss function between the predicted label of the production sample image y and the class label s, L c (y) represents the loss function between the predicted label of the production sample image y and the c-class label, L d (y) represents the loss function between the predicted label of the production sample image y and the d-class label: Wherein, p(h|y) represents the relationship between the production sample image y and the label h. When the label of the production sample image y is the same as the label h, p(h|y)=1. When the label of the production sample image y is different from the label h, p(h|y)=0. S represents the set of s-type labels. When the label corresponding to the final text feature is label h, the marginal probability between the production sample image y and the final text feature, p(j|y) represents the relationship between the production sample image y and label j, when the label of the production sample image y is the same as label j, p(j|y)=1, when the label of the production sample image y is different from label j, p(j|y)=0, C represents the set of c-type labels, When the label corresponding to the final text feature is label j, the marginal probability between the production sample image y and the final text feature, p(m|y) represents the relationship between the production sample image y and label m, when the label of the production sample image y is the same as label m, p(m|y)=1, when the label of the production sample image y is different from label m, p(m|y)=0, D represents the set of d-type labels, Indicates the marginal probability between the production sample image y and the final text feature when the label corresponding to the final text feature is label m; Determine whether the loss function value of the multi-task classification model is less than a preset loss function value; If yes, the multi-task classification model is used as the optimized multi-task classification model; Otherwise, adjust the parameters in the multi-task classification model, return the final image features and all final text features based on the production sample image, and use the multi-task classification model to obtain the marginal probability between the production sample image and each of the final text features.

7. The pellet production condition detection method according to claim 1, characterized in that: The method of obtaining the working condition detection result of the pellet production that needs to be tested based on all final marginal probabilities includes: Select the final marginal probability with the largest value from all final marginal probabilities as the target marginal probability; The label corresponding to the target marginal probability is used as the operating condition detection result of the pellet production that needs to be detected.

8. A pelletizing production condition detection device, characterized in that: include: A labeling module, which obtains a plurality of production working condition images of pellet ore, and respectively performs text labeling on each of the production working condition images to obtain a label of the production working condition image; A feature extraction module, using an image feature extraction model to extract features from each of the production working condition images to obtain image features of the production working condition images; An encoding module, using a text encoder to encode the label of each of the production working condition images respectively, to obtain text features of the production working condition images; A first optimization module optimizes the image feature extraction model and the text encoder according to the image features and text features of all production working condition images to obtain an optimized image feature extraction model and an optimized text encoder; A first acquisition module, using the optimized text encoder to acquire final text features of each of the production working condition images; A second acquisition module, using the optimized image feature extraction model to acquire final image features of the production sample image; The production sample image is a production image of pellets with actual labels; A marginal probability acquisition module, based on the final image features and all the final text features of the production sample image, using a multi-task classification model to acquire the marginal probability between the production sample image and each of the final text features; A second optimization module optimizes the multi-task classification model according to all marginal probabilities to obtain an optimized multi-task classification model, and uses the optimized multi-task classification model to obtain a final marginal probability between a production image of pellet production that needs to be tested for working conditions and each of the final text features; The detection module obtains the working condition detection result of the pellet production that needs to be detected based on all final marginal probabilities.

9. A terminal device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the pelletizing production condition detection method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method for detecting the pelletizing production conditions as described in any one of claims 1 to 7 is implemented.